Cracow Grid Workshop, October 27 – 29, 2003

Institute of Computer

Science AGH

Design of Distributed Grid Workflow Composition System

Marian Bubak, Tomasz Gubała, Maciej Malawski, Katarzyna Rycerz

Institute of Computer Science & ACC CYFRONET - AGH Kraków, Poland

Cracow Grid Workshop, October 27 – 29, 2003

Institute of Computer

Science AGH

Outline

Motivation - workflow composition problem

First prototype – AFC Architecture of the Flow Composition

Sytem Distributed registry subsystem Agent-based composition subsystem Application of ontology Summary

Cracow Grid Workshop, October 27 – 29, 2003

Institute of Computer

Science AGH

Grid Application Workflow Simple definition of a Grid application

A Grid application is a combination of cooperating Grid resources (application elements) joined together to achieve particular functionality

Workflow of Grid application

The distributed computation is performed by application elements using a remote procedure call facility for co-ordination (data flow though means of procedure parameters passing)

Cracow Grid Workshop, October 27 – 29, 2003

Institute of Computer

Science AGH

Flow Composition Problem To execute an application on the Grid the

application workflow description is needed

Workflow description should contain:• list of application elements (components, services,

etc.)• list of element connections (control or data links)• technology and framework specific deployment

data

Problems:• lack of important information• frequent changes of environment• problem not well understood (Grid prototyping)

Cracow Grid Workshop, October 27 – 29, 2003

Institute of Computer

Science AGH

First Prototype - AFC

Application Flow Composer system

Two main modules Designed for

Common Component Architecture (CCA)

Monolithic registry as GT3 service

Implemented in XCAT Framework with Java

Cracow Grid Workshop, October 27 – 29, 2003

Institute of Computer

Science AGH

Advantages and Drawbacks of the AFC

Clear, modular design

Portable (XML + Java) Two Grid

technologies used Successful as a

feasibility study Encouraging

efficiency test results

Consolidated, centralized registry

Non fault-tolerant Simple decision

module (no semantic data used)

High communication and registry search overhead

Cracow Grid Workshop, October 27 – 29, 2003

Institute of Computer

Science AGH

Conclusions from Prototype

To exploit advantages• stay with the clear decomposition model• use portable platform and languages• use popular Grid development technology (OGSA)

To eliminate weaknesses• introduce full decentralization of data and control• add fault tolerance to overcome Grid failures• apply ontology for semantic service description• implement migration mechanism to reduce distant

communication overhead

Cracow Grid Workshop, October 27 – 29, 2003

Institute of Computer

Science AGH

Proposed Architecture of the Flow Composition System

Grid Site

Service Provider Registry Administrator External Registry User

User (Portal)

Grid Site

Service registrationFull service entry data

Registry up/downPolicy management

Workflow composition requestSearch hints (system learning)

Results evaluation (system learning)

DedicatedAFC connection

Requests transferWorkflows transferCurrent system topology info

Traditional syntacticservice discovery

Inter-peer operationsEntries custody transfer

Entries ownership transferContent coherency replication

Specialization info (publication)

Grid Site

Peer migrationor spawning

Directservicesbrowsing

Few possible levels ofComposer-Registry coupling

Registry NodeRegistry Node

Registry Node

Composer Agent

Composer Agent

Composer Agent

Cracow Grid Workshop, October 27 – 29, 2003

Institute of Computer

Science AGH

Composer and Registry

Composer Agent and Registry Node can be• one program entity exposing externally

two different behaviors through various interfaces

• distinct entities with registry independence and agents as registry plug-ins

• complete separated bodies, communicating through standardized protocol over Internet link

Cracow Grid Workshop, October 27 – 29, 2003

Institute of Computer

Science AGH

Stages of Workflow Composition

User defines initial conditions in Initial Workflow Description

Flow Composer divides IWD into smaller subproblems

Each simple problem is mapped to one existing service

Information about services is achieved from Service Registry

Ready mini-workflows are concatenated into one The workflow is returned to user as Final

Workflow Document (FWD)

Cracow Grid Workshop, October 27 – 29, 2003

Institute of Computer

Science AGH

Registry Node

Interfaces:• management for admin• register for service

provider• lookup for users• dedicated lookup for FC

Internals:• Service entries table• Peer registry nodes table• Data replication

subsystem• Specialization

description• Query routing table

AFC-dedicated lookup

looku

p

regis

ter

management

Registry Node

Service EntryService Entry

Node Entry

ServiceCategoriesDescription

InterfaceSyntactic

Information

CapabilitiesSemantic

Description

ServiceName and

Localization

ReplicationInformation

Query RoutingRules

Grid ServiceInformationDoc.(GSID)

SpecializationDescription

Cracow Grid Workshop, October 27 – 29, 2003

Institute of Computer

Science AGH

Grid Service Information Document

Service identification and localization data Service category (application of which

domain do use this service frequently) Interface syntactic information (list of

methods, signatures, implemented ports) Semantics description (service

capabilities, method results meaning, pre-conditions and post-conditions of method invocation)

Cracow Grid Workshop, October 27 – 29, 2003

Institute of Computer

Science AGH

Data Replication

Whole structure consists of many registries connected for inter-registry communication

One registry includes a few nodes Every node data is replicated partially in

every other node Full recovery in case of single node failure

(all the data may be acquired by means of existing nodes cooperation)

Cracow Grid Workshop, October 27 – 29, 2003

Institute of Computer

Science AGH

Composer Agent

Interfaces:• Interface with registry• Flow composition for

users• Inter-agent interface

Internals:• Matchmaker based on

ontology inference engine

• Problem decomposition and workflow composition mechanism

• Migration subsystem

AFC

registry interface

flow composition

Composer Agent

SemanticMatchmaker

RequestDecomposer

WorkflowComposer

MigrationSubsystem

Cracow Grid Workshop, October 27 – 29, 2003

Institute of Computer

Science AGH

Mobility of Agent

Purpose: to reduce the registry querying time by localization of the communication

Agent behaves as a stateless entity but has its internal state - this implies relocation instead simple re-spawn

Migration involves historical data and rules set marshalling and transfer

Problem of widespread platform

Cracow Grid Workshop, October 27 – 29, 2003

Institute of Computer

Science AGH

Application of Ontology

Semantic description of services in GSID Ontology reasoning to provide accurate

matchmaking facility Semantic description of registry

specialization Specification of rules for automatic

complex problem decomposition

Cracow Grid Workshop, October 27 – 29, 2003

Institute of Computer

Science AGH

Summary – Expected Improvements

Portable for many architectures (highly portable languages)

High scalability (evolves with the Grid environment) Lightweight underlying platform (deploying

automation) Fault tolerance support (data distribution and

replication) Adaptable to changing environment (dynamic

topology) Faster request handling (work distributed to many

nodes) More accurate decisions (self-learning engines,

semantic service and registry description)